Suggest Loop

suggest-loop

Generation exercise, not template fill. Turn a repo's verification signal into /loop prompts a "vibe coder" can paste as-is — someone who can't author a precise loop prompt, so the repo proposes it and they just approve or tweak. The failure mode: a confident-sounding loop with no measurable "done".

Read this first — the load-bearing constraint

Do not reason about /loop from memory. /loop, the Monitor tool, and /goal post-date the model's training cutoff; a session reconstructing them from memory gets them wrong (that's why this skill exists). Read [references/loop-mechanics.md](references/loop-mechanics.md) before emitting any suggestion — every claim about how /loop behaves comes from there, not priors.

Recipe

1. Find the hard gate. Read the target repo's CLAUDE.md, package.json

scripts, CI config, test setup — whatever documents how work is checked. You're looking for a machine signal: an exit code, a measured assertion, an emptying queue. If the repo documents no verification signal, stop and say so — suggest the repo document one first; do not fabricate a gate. A suggestion is only as good as the documented oracle.

1. Match the gate to real work. A failing suite, a backlog item, a scoped

feature, a refactor that must stay green. The loop has to be about something in this repo, not a generic shape.

1. Emit an inline /loop with the stop condition baked in. "…stop when the

suite is green twice in a row", "…stop when the work-list is empty". Inline is primary. Reach for .claude/loop.md only for a loop run often; reach for /goal only when "done" is judgment-heavy (see the mechanics file). Default to a plain inline /loop.

1. Mark the human/taste-gate explicitly. Autonomy follows the oracle: where

correctness is a fact about the world (an exit code, a hash, a measured peak) the loop checks itself; where "correct" lives in the user's taste (does it sound good, does the UI feel right) a human gates every step. Most repos are a split — scope the loop to the machine-measurable half and hand the taste half back ("the by-ear check stays yours"). Naming that seam is the point, not a footnote.

Anti-patterns — refuse these, don't dress them up

• Vague termination ("fix until no more bugs", "keep improving it") — no

measurable done, the loop never knows it's finished. Rewrite to a hard signal or decline.

• Taste/ear-gated work (does it sound right, does the UI feel right) —

not loopable autonomously; the oracle is a human sense. Say so and leave it human-gated.

• No documented signal — a suggestion would be guessed. Recommend

documenting the gate instead of inventing one.

When you refuse, that is the deliverable — a clear "this isn't loopable, and here's the one thing that would make it loopable" beats a plausible-looking loop that runs to nowhere.

Output shape

Per loop, a short scannable block: the paste-ready /loop line, the hard signal it terminates on, and the taste-gate line.

> Loop 1 — keep the suite green (the clean first loop) > `` > /loop run pnpm test, fix the single top failure, re-run; commit each green > step; stop when pnpm test exits 0 twice in a row > ` > - **Hard signal:** exit code of pnpm test`. > - Stays yours: whether the result feels right — animation smoothness, > copy tone, visual polish — the loop never touches those.

A queue-loop must carry its own queue. Derive the work-list inside the loop line from source files ("build the list of X from a.ts, then drain it"), not from a STATUS/backlog doc that may be stale or a list the user never gave. If the queue can't self-derive, the loop isn't paste-ready — mark the gap or send it to the refusals.

Scope

This skill writes prompt text; it does not run anything. Suggest 2–3 good loops, name what can't be looped, and stop — don't force a loop for every corner.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ikkeseb
• Source: ikkeseb/skills
• License: MIT

Install and usage instructions live in the source repository linked above.